17
$\begingroup$

Harbour porpoises produce distinct narrow band high frequency clicks centered around 130 kHz. Typically these are usually easy to identify by a manual analyst and should be relatively easy to automatically identify because there are few confounding sounds at those high frequencies. However, as with all automated analysis, it's not quite that simple; off-axis clicks, reflections causing distortion, noise from boat engines and low SNR clicks can be difficult for automated classifiers to cope with, leading to false positive and false negatives.

There are huge manually annotated datasets of porpoise clicks collected over decades - so a generic automated classifier seems like a prime candidate for a CNN or other deep learning approach. Does anyone know of or is working on a porpoise click classifier based on deep learning methods?

Improve this question
$\endgroup$

3 Answers 3

Reset to default
9
$\begingroup$

Dave Mellinger and I did some work on this; it's been presented at conferences but not yet published. Below is the abstract for ASA and a link the final report to the funding agency.

Conference abstract: Mellinger, D. K., & Fregosi, S. (2021). FindPorpoises: Deep learning for detection of harbor porpoise echolocation clicks. The Journal of the Acoustical Society of America, 149(4), A39-A40.

Report: https://oera.ca/research/improved-analysis-harbour-porpoise-sounds

I primarily did the manual analysis and reporting aspects of this project so I can't speak much to the deep-learning part, but Dave would be happy to discuss, I'm sure!

Improve this answer
$\endgroup$
1
  • $\begingroup$ Great, thanks very much. Do you now if the model is open source? $\endgroup$
    – user213
    Commented Jul 21, 2022 at 8:41
7
$\begingroup$

As Dr. Fregosi mentioned, we made a deep-learning classifier for porpoise clicks. But Daniel De Leon, a recently-graduated student at Oregon State U, did better work with a more complete version. His masters thesis about it (Automatic Classification of Ultrasonic Harbor Porpoise Clicks in a Varying Noise Environment) is here: https://ir.library.oregonstate.edu/concern/graduate_thesis_or_dissertations/9k41zn95b

I'll tell him you're interested and he may contact you. I'm guessing he'd be happy to share code or expertise. It might not be exactly what you want, since it was built for a specific noise environment (Minas Passage, Bay of Fundy), but hopefully it will help. I think his code is on Github but I'm not certain.

Dave Mellinger

Improve this answer
$\endgroup$
1
  • $\begingroup$ Great, thanks very much. If Daniel is interested would be great to hear some more about his work and have a look at the code. $\endgroup$
    – user213
    Commented Jul 21, 2022 at 18:38
4
$\begingroup$

I'll also mention that in high-noise environments like the Bay of Fundy, it's not always easy to tell a harbor porpoise click from background noise. Sometimes you get things in the right frequency range for which it's quite hard to tell whether or not they are porpoise clicks. The spectrum is right, and maybe there's another click with the right time spacing, but that's it. Was that a porpoise or just random noise? You can spend a good chunk of time on a single click and still not be sure.

Granted it's an unusual noise environment, with extremely fast currents (10 knots!) and gravel clacking around and sand hitting the hydrophone.

Improve this answer
$\endgroup$
1
  • 3
    $\begingroup$ Hi @Dave Mellinger, would you consider adding this comment into the original answer? It might help reduce confusion for readers and also help the answer get more consolidated points+ attention. Long and detailed answers are welcome on SE! $\endgroup$
    – Thejasvi
    Commented Jul 22, 2022 at 8:03